Industrial process fault detection method based on mutual k local neighborhood standardization analysis

By adopting the mutual k local neighborhood standardization method in the multimodal industrial process, redefining the mean and standard deviation of the data, the pseudo-nearest neighbor problem in high-dimensional and multimodal data processing is solved, and the fault detection performance of PCA is improved. The normal distribution conversion of data and the accuracy of fault detection are achieved.

CN119066407APending Publication Date: 2024-12-03SHENYANG INSTITUTE OF CHEMICAL TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411020714.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

In the multimodal industrial process, it is difficult for the prior art to effectively process high-dimensional, multimodal data and contain a large number of fault points, resulting in a degradation of PCA fault detection performance, especially when outliers or noise exist, pseudo-nearest neighbors during data standardization process The impact is difficult to eliminate.

Method used

Using a method based on mutual k local neighborhood standardization, the mutual k nearest neighbor set of each data point is calculated, and the mean and standard deviation of the data are redefined to ensure that the data is consistent with the normal distribution after standardization, thereby improving PCA in the multimodal industrial process. fault detection performance in.

Benefits of technology

Effectively convert multimodal data into single modal data, reduce the influence of pseudo-nearest neighbors, improve the accuracy and performance of PCA in complex industrial processes, and ensure the reliability of detection results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure SMS_9
    Figure SMS_9
  • Figure SMS_11
    Figure SMS_11
  • Figure SMS_19
    Figure SMS_19
Patent Text Reader

Abstract

The invention discloses an industrial process fault detection method based on mutual k local neighborhood standardization analysis, and relates to an industrial process fault detection method. The method comprises the following steps: firstly, collecting multi-modal data in an industrial production process; under the condition that the k value is fixed, screening k neighbors of random sample points in the training sample set through Euclidean distance; screening the first k neighbors of the sample points in the neighbor set according to the Euclidean distance; judging whether the first k neighbors of the neighbor sample points contain the original training sample points or not, and judging the number of the k neighbors of each sample point in the training set; and establishing a related principal component analysis process detection model, and starting fault monitoring on new data. According to the invention, the monitoring performance of principal component analysis in multi-modal fault detection is improved. Experimental results show that the method can effectively perform multi-modal industrial process fault detection, plays a good role in monitoring the operation state of the industrial process, improves the accuracy of industrial process fault detection, and reduces the false detection rate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an industrial fault detection method, and in particular to an industrial process fault detection method based on mutual k local neighborhood standardized analysis. Background Art

[0002] With the increasing complexity of production processes and the continuous expansion of production scale, the safety risks in the production process are also increasing. Therefore, data-driven fault detection methods have attracted widespread attention from researchers. Among them, multivariate statistical process monitoring methods represented by principal component analysis (PCA) have been widely used in a variety of industrial production processes. If the industrial process data has only one working mode, then only by extracting data features and reducing dimensions through PCA, identifying data that deviates from the data center as fault samples can achieve good fault detection performance. Therefore, PCA requires that the data has only one stable mode, but industrial processes often contain multiple working states, so most industrial process data are multimodal, which greatly reduces the monitoring efficiency of PCA.

[0003] For the monitoring of multi-mode processes, the main improvement methods can be divided into three categories: multi-model / decision function scheme, multi-model / Bayesian function scheme and single model scheme. Among them, the multi-model / decision tree method can improve the accuracy and robustness of detection, but at the same time it also increases the complexity of the detection system to a certain extent; and for the multi-model / Bayesian function method, it also improves the robustness and reliability of the detection system, but this method, like the traditional PCA monitoring method, requires certain prior knowledge. Therefore, the simple and easy-to-implement single model scheme has become the most widely studied fault detection scheme. This scheme mainly involves steps such as data preprocessing, model selection, and training of the constructed model. The result of data preprocessing directly affects the subsequent model's learning of data features and the fault detection effect of the overall model. Most of the commonly used data preprocessing methods are based on the z-score method. Although this method is simple and direct, if the detected data is more complex or the pattern is not obvious in industrial situations, the detection accuracy of this method will be greatly reduced.

[0004] In order to ensure that PCA has higher monitoring performance in multimodal industrial processes and can effectively eliminate the multimodal characteristics of data, some data preprocessing methods based on local modeling strategies have been proposed, such as: a data preprocessing strategy based on local neighborhood standardization (LNS). Compared with the traditional z-score, the data preprocessing strategy of local neighborhood standardization uses the local neighborhood mean and standard deviation for standardization instead of using a fixed global mean and standard deviation. However, when encountering a data set containing a large number of outliers or noise, the nearest neighbor found in the process of LNS data standardization may not be an actual or reliable neighborhood, but a pseudo-neighborhood, which will reduce the performance of process monitoring. Therefore, it is very necessary to convert the multimodal non-Gaussian features in the multimodal data of complex industrial processes into a Gaussian distribution that approximately satisfies a single mode, while minimizing the influence of pseudo-neighbors in the data set as much as possible, for the fault detection performance of PCA in complex industrial processes. Summary of the invention

[0005] The present invention proposes an industrial process fault detection method based on mutual k local neighborhood normalization analysis. The present invention realizes the conversion of multimodal data into single modal data when the data is high-dimensional, multimodal and has a large number of fault points, and converts non-Gaussian distributed process variables into approximately Gaussian distribution, thereby improving the fault detection performance of PCA in multimodal industrial processes. The method solves the complex data characteristics of industrial process data such as high dimension, multimodality, and a large number of fault points, as well as the fault detection problem.

[0006] The technical solution adopted by the present invention is:

[0007] The present invention proposes an industrial process fault detection method based on mutual k local neighborhood normalization analysis. The method is based on a mutual k local neighborhood normalization algorithm and applies the algorithm to industrial process fault detection. The steps are as follows:

[0008] Step 1: Assume X∈R n×m is a standardized multimodal data matrix, where n is the number of samples and m is the number of variables. Calculate each training sample point x i The Euclidean distance between all sample points and arranged in ascending order, we get formula (1):

[0009]

[0010] Among them, N(x i ) is the data point x in the training data set i The distances from all training data are sorted in ascending order. is the data point x iThe jth nearest neighbor (1≤j≤k) of x is taken to form the first k nearest neighbors of x. i The first k nearest neighbor set of , we get formula (2):

[0011]

[0012] Step 2: Based on formula (2), redefine the set N k (x i ),make Calculate x separately i Each k nearest neighbor point x j The Euclidean distance between the remaining data points in the entire training set is obtained by repeating the process of formula (2) for each x j The first k nearest neighbor set of is expressed as:

[0013]

[0014] Step 3: According to the relevant definition of mutual k-nearest neighbors, if x i The first k nearest neighbors of the data point x j , and x j There is a data point x in the first k nearest neighbors of i , then these two data points are called mutual k-nearest neighbors, and can be expressed as formula (4) and formula (5):

[0015] M k (x i )={x i ∈N k (x j )^x j ∈N k (x i )} (4)

[0016]

[0017] Among them, N k (x j ) is x j The k nearest neighbor set, N k (x i ) is x i The k nearest neighbor set of Represents x i The nth mutual k-nearest neighbor of k (x i ) is an empty set, then prove that x i There are no mutual k-nearest neighbors, that is, at this time the data point x i The k nearest neighbors of x are all pseudo-nearest neighbors. i Belongs to outliers or border points.

[0018] Step 4: When the number of mutual k-nearest neighbors is 0 or 1, choose to use x when calculating the mean and standard deviation of the data. i The number of data points N in the neighborhood of the first nearest neighbor replaces the number of its k-nearest neighbors K; the formula for the data mean and standard deviation can be changed to formula (6), and the training data point x i Neighborhood set of the first nearest neighbor It can be expressed by formula (7):

[0019]

[0020]

[0021] The standardization formula is now updated to the first equation in equation (8). When the number of mutual k-nearest neighbors is greater than 1, since the smaller the standard deviation used in the data standardization process, the better, the standard deviation of the first nearest neighbor set is still used here, which also reduces the amount of model calculation to a certain extent. In order to ensure that the standardized data follows a normal distribution with a mean of 0, this paper selects x i The mutual k-nearest neighbor set mean replaces the first nearest neighbor set mean, and the standardized formula is improved accordingly. At this time, the improved standardized formula is the second equation in formula (8).

[0022] The standardized formula obtained through the above analysis is shown in formula (8):

[0023]

[0024] In the formula is the standardized data, and are the neighborhood mean and standard deviation of the first nearest neighbor; m(M k (x i )) is x i The mean of the first K nearest neighbor sample sets, K represents x i The number of mutual k-nearest neighbors.

[0025] A PCA process monitoring model is established based on the related information of the mutual k-nearest neighbors.

[0026] Step 5: Perform an orthogonal transformation on the standardized data and the transformation matrix is The data set is decomposed into the principal component subspace and residual space as shown in formula (9):

[0027] X=TP T +X=TP T +TPT (9)

[0028] Where T∈R m×l is the score of X in the l-dimensional principal component subspace, Residual, P∈R n×l is the loading matrix of the principal component subspace.

[0029] To detect the process, two traditional statistics T 2 and SPE with their control limits and SPE α According to the following formula:

[0030] T 2 =x new PΛ -1 P T x new (10)

[0031]

[0032]

[0033]

[0034] where θ r and h0 satisfy equation (14):

[0035]

[0036] Among them, Λ in formula (10) -1 is the diagonal matrix of l largest eigenvalues. In formula (11), I is the unit matrix. In formula (12), F α (l,nl) represents the critical value of the F distribution when the significance level is α and the degrees of freedom are l and nl.

[0037] Step 6: The fault detection phase is divided into the offline modeling phase and the online modeling phase. The process is as follows:

[0038] Offline modeling steps:

[0039] 1) Use normal data of different modes as training dataset X∈R n×m ;

[0040] 2) For each training sample point x i , calculate the distance between it and other sample points in the training data set according to the Euclidean distance formula;

[0041] And sort them in ascending order, and filter out the training sample points x according to the sorting results i The first k nearest neighbor point set N k (x i )

[0042] make

[0043] 3) For the selected neighbor sample set N k (x i ) for each sample point Repeat step 2 and filter out The first k nearest neighbors of

[0044] Sample point set

[0045] 4) According to formula (4), select the training sample point x i The mutual k sample set M k (x i );

[0046] 5) If M k (x i ) is an empty set or contains only one training sample point, then calculate the training sample point x i The mean of the first nearest neighbor of

[0047] and standard deviation, and use equation 1 in equation (8) to calculate x i To standardize;

[0048] 6) If M k (x i ) contains K training sample points, and K>1, then calculate the training sample point x i The mean of the mutual k-nearest neighbor sample set of x is calculated by using equation 2 in equation (8) i To standardize;

[0049] 7) For each training sample point after standardization The principal component analysis method is used to extract the principal components and perform dimension reduction;

[0050] 8) Calculate the relevant T 2 and control limits of SPE.

[0051] The online modeling phase mainly includes the following five steps:

[0052] 1) When a new sample point x enters the data set new , calculate the Euclidean distance between the sample point and all sample points in the training data set, and sort them in ascending order, and filter out the sample point x according to the sorting result new The first k nearest neighbors N k (x new );

[0053] 2) In the modeling stage, the new samples are standardized using steps 3, 4, 5, and 6 in offline modeling.

[0054] 3) Project the standardized new samples into the main residual space;

[0055] 4) Use equations (10) and (11) to calculate new samples The statistic T 2 and SPE;

[0056] 5) Comparison The statistic T 2 and control limits of SPE. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of the Tennessee Eastman process structure;

[0058] Figure 2 It is a flow chart of the principal component analysis based on mutual k local neighborhood normalization of the present invention;

[0059] Figure 3 Time series diagram of Tennessee Eastman multimodal data for traditional data normalization methods;

[0060] Figure 4 Distribution plots of Tennessee Eastman multimodal data for traditional data normalization methods;

[0061] Figure 5 A time series diagram of the present invention processing Tennessee Eastman multimodal data based on mutual k local neighborhood normalization;

[0062] Figure 6 The distribution diagram of the Tennessee Eastman multimodal data based on the mutual k local neighborhood normalization of the present invention;

[0063] Figure 7 The T of mode 1 fault 18 based on the principal component analysis of mutual k local neighborhood normalization of the present invention is 2 Monitoring results diagram;

[0064] Figure 8 This is a diagram of the SPE monitoring result of mode 1 fault 18 based on the principal component analysis of mutual k local neighborhood normalization of the present invention;

[0065] Fig. 9 The principal component analysis based on the k-closest local domain normalization of the present invention is used to analyze the T of mode 3 fault 13. 2 Monitoring results diagram;

[0066] Fig.10 This is a diagram of the SPE monitoring results of mode 3 fault 13 based on the principal component analysis of mutual k local neighborhood normalization of the present invention. DETAILED DESCRIPTION

[0067] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0068] The present invention takes the Tennessee Eastman process as an example. The schematic diagram of the Tennessee Eastman process is as follows: Figure 1 As shown in the figure, the fault detection flowchart based on mutual k local neighborhood normalization is as follows Figure 2 As shown in the figure, the traditional data standardization method processes multimodal data result information as follows Figure 3 , Figure 4 As shown in the figure, the multimodal data result information based on mutual k local neighborhood normalization is as follows Figure 5 , Figure 6 As shown in the figure, the fault detection results of the Tennessee Eastman process based on principal component analysis based on mutual k local neighborhood normalization are as follows: Figures 7 to 10 shown.

[0069] Specific experimental steps:

[0070] Step 1: Collect sample data from industrial processes. The training set and test set in this paper are collected based on the revised version of Bathelt. The Tennessee Eastman process mainly consists of five unit operations, including reactors, condensers, circulating compressors, vapor-liquid separators, and stripping towers. The initial version of TEP simulation involves 12 manipulated variables and 41 measured variables. In 2015, Bathelt revised the original TE model and established a new TEP model using Simulink in MATLAB, adding 8 measured variables, 24 component variables, and 8 new process disturbances. During the simulation, 9 manipulated variables and 22 continuous process variables were selected for monitoring. This paper selects two modes, mode 1 and mode 3, to simulate the TEP multi-mode process. Generate two modes of normal data sets and 20 fault data sets, where each mode of normal data set contains 500 samples, a total of 1000 samples to form a training set, and each fault test data set contains 1000 samples, with a sampling time of 0.05h. All faults are introduced from the 201st sample until the end of the test data set. Calculate each training sample point x i The Euclidean distance between all sample points and arranged in ascending order, we get formula (15):

[0071]

[0072] Among them, N(x i ) is the data point x in the training data set i The distance between the training data and the training data is in ascending order. In the present invention, k is set to 15. The purpose of this improvement is to better reduce the influence of outliers on local information in the process of finding mutual k-nearest neighbors and optimize the effect of standardization of the training data set. is the data point x i The jth nearest neighbor (1≤j≤k) of x is taken to form the first k nearest neighbors of x. iThe first k nearest neighbor set of , we get formula (16):

[0073]

[0074] Step 2: Based on formula (16), redefine the set N k (x i ),make Calculate x separately i Each k nearest neighbor point x j The Euclidean distance between the remaining data points in the entire training set is obtained by repeating the process of formula (16) for each x j The first k nearest neighbor set of is expressed as:

[0075]

[0076] Step 3: According to the relevant definition of mutual k-nearest neighbors, if x i The first k nearest neighbors of the data point x j , and x j There is a data point x in the first k nearest neighbors of i , then these two data points are called mutual k-nearest neighbors, and can be expressed as formula (18) and formula (19):

[0077] M k (x i )={x i ∈N k (x j )^x j ∈N k (x i )} (18)

[0078]

[0079] Among them, N k (x j ) is x j The k nearest neighbor set, N k (x i ) is x i The k nearest neighbor set of Represents x i The nth mutual k-nearest neighbor of k (x i ) is an empty set, then prove that x i There are no mutual k-nearest neighbors, that is, at this time the data point x i The k nearest neighbors of x are all pseudo-nearest neighbors. i Belongs to outliers or border points.

[0080] Step 4: When the number of mutual k-nearest neighbors is 0 or 1, choose to use x when calculating the mean and standard deviation of the data. i The number of data points N in the neighborhood of the first nearest neighbor replaces the number of its k-nearest neighbors K; the formula for the data mean and standard deviation can be changed to formula (20), and the training data point x i Neighborhood set of the first nearest neighbor It can be expressed by formula (21):

[0081]

[0082]

[0083] The standardization formula is now updated to the first equation in equation (8). When the number of mutual k-nearest neighbors is greater than 1, since the smaller the standard deviation used in the data standardization process, the better, the standard deviation of the first nearest neighbor set is still used here, which also reduces the amount of model calculation to a certain extent. In order to ensure that the standardized data follows a normal distribution with a mean of 0, this paper selects x i The mutual k-nearest neighbor set mean replaces the first nearest neighbor set mean, and the standardized formula is improved accordingly. At this time, the improved standardized formula is the second equation in formula (22).

[0084] The standardized formula obtained through the above analysis is shown in formula (22):

[0085]

[0086] In the formula is the standardized data, and are the neighborhood mean and standard deviation of the first nearest neighbor; m(M k (x i )) is x i The mean of the first K nearest neighbor sample sets, K represents x i The number of mutual k-nearest neighbors.

[0087] A PCA process monitoring model is established based on the related information of the mutual k-nearest neighbors.

[0088] Step 5: Perform an orthogonal transformation on the standardized data and the transformation matrix is Decompose the data set into principal component subspace and residual space as shown in formula (23),

[0089] X=TP T +X=TP T +TP T (23) where T∈R m×l is the score of X in the l-dimensional principal component subspace, X represents the residual, P∈R n×lis the loading matrix of the principal component subspace.

[0090] To detect the process, two traditional statistics T 2 and SPE with their control limits and SPE α According to the following formula:

[0091] T 2 =x new PΛ -1 P T x new (twenty four)

[0092]

[0093]

[0094]

[0095] where θ r and h0 satisfy equation (28)

[0096]

[0097] Among them, Λ in formula (24) -1 is the diagonal matrix of l largest eigenvalues. In formula (25), I is the unit matrix. In formula (26), F α (l,nl) represents the critical value of the F distribution when the significance level is α and the degrees of freedom are l and nl.

[0098] Step 6: The fault detection phase is divided into the offline modeling phase and the online modeling phase. The process is as follows:

[0099] Offline modeling steps:

[0100] 9) Use normal data of different modes as training data set X∈R n×m ;

[0101] 10) For each training sample point x i , calculate the distance between it and other sample points in the training data set according to the Euclidean distance formula;

[0102] And sort them in ascending order, and filter out the training sample points x according to the sorting results i The first k nearest neighbor point set N k (x i )

[0103] make

[0104] 11) For the selected neighbor sample set Nk (x i ) for each sample point Repeat step 2 and filter out The first k nearest neighbors of

[0105] Sample point set

[0106] 12) According to formula (4), filter out the training sample points x i The mutual k sample set M k (x i );

[0107] 13) If M k (x i ) is an empty set or contains only one training sample point, then calculate the training sample point x i The mean of the first nearest neighbor of

[0108] and standard deviation, and use equation 1 in equation (8) to calculate x i To standardize;

[0109] 14) If M k (x i ) contains K training sample points, and K>1, then calculate the training sample point x i The mean of the mutual k-nearest neighbor sample set and the standard deviation of the k-nearest neighbors, and use equation 2 in equation (8) to calculate x i To standardize;

[0110] 15) For each training sample point after standardization The principal component analysis method is used to extract the principal components and perform dimension reduction;

[0111] 16) Calculate the relevant T 2 and control limits of SPE.

[0112] The online modeling phase mainly includes the following five steps:

[0113] 6) When a new sample point x enters the data set new , calculate the Euclidean distance between the sample point and all sample points in the training data set, and sort them in ascending order, and filter out the sample point x according to the sorting result new The first k nearest neighbors N k (x new );

[0114] 7) In the modeling stage, the new samples are standardized using steps 3, 4, 5, and 6 in offline modeling.

[0115] 8) Project the standardized new samples into the main residual space;

[0116] 9) Use equations (10) and (11) to calculate the new sample The statistic T 2 and SPE;

[0117] 10) Comparison The statistic T 2 and control limits of SPE.

[0118] Analysis of simulation experiment results:

[0119] from Figure 3 It can be seen that after the multimodal data is processed by the traditional data standardization method, its multimodal characteristics still exist, and Figure 4 It can be seen that after multimodal data is processed by traditional data standardization methods, its data does not satisfy the normal distribution. Figure 5 It can be seen that after the multimodal data is processed by the mutual k local neighborhood normalization method, its multimodal features are erased and Figure 6 It can be seen that after the multimodal data is processed by the mutual k local neighborhood normalization method, its data satisfies the normal distribution.

[0120] In this embodiment, the two monitoring statistics T under mode 1 fault 18 are 2 and SPE are 82.13% and 73.25% respectively. The two monitoring statistics T under mode 3 fault 13 2 The SPE and SPE are 94.63% and 89.5% respectively. The process monitoring method of the invention shows good fault detection performance. Next, the fault detection performance of the invention is analyzed. The present invention uses PCA to perform fault monitoring and analysis on industrial process data. The detection results are as follows: Figures 7 to 10 As shown in the figure, the red dotted line represents the control limit of the monitoring statistic. The present invention takes a control limit of 99%. It can be clearly seen that in the first 201 normal samples, no false positives are generated, or the false positive rate is very small, which shows that it is reasonable for the present invention to set the control limit to 99%. Figure 7 As shown in , in the original data, the first 200 samples are normal data, and the detection results are consistent with expectations, without any false detection; secondly, Figure 7 , Figure 8 As shown in Figure 2, although the detection results are only 82.13% and 73.25%, they are much better than traditional PCA. Fig. 9 , Fig.10 As shown in Figure 2, the test data set introduces faults at 201 sample points. The test results show that although the time of detecting the fault is delayed compared with the expected time of detecting the fault, the T2 statistic reaches 94.63%, which greatly improves the detection performance of PCA. Finally, from Figures 7 to 10It can be seen that after the mutual k local neighborhood normalization process, the model can correctly identify fault data and normal data, thereby completing the fault detection task.

[0121] In summary, the feasibility of MKLNS-PCA in multimodal fault detection and the effectiveness of the MKLNS method in converting multimodal data into a single normally distributed data are demonstrated.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. Industrial process fault detection method based on mutual k local neighborhood normalization analysis, It is characterized in that The method comprises the following steps: Firstly, an algorithm based on mutual k local neighborhood normalization is proposed and applied to industrial process fault detection. The steps are as follows: Step 1: Assume X∈R n×m is a standardized multimodal data matrix, where n is the number of samples and m is the number of variables; Calculate each training sample point x i The Euclidean distance between all sample points and arranged in ascending order, we get formula (1): Among them, N(x i ) is the data point x in the training data set i The distances from all training data are sorted in ascending order. is the data point x i The jth nearest neighbor (1≤j≤k) of x is taken to form the first k nearest neighbors of x. i The first k nearest neighbor set of , we get formula (2): Step 2: Based on formula (2), redefine the set N k (x i ),make Calculate x separately i Each k nearest neighbor point x j The Euclidean distance between the remaining data points in the entire training set is obtained by repeating the process of formula (2) for each x j The first k nearest neighbor set of is expressed as: Step 3: According to the relevant definition of mutual k-nearest neighbors, if x i The first k nearest neighbors of the data point x j , and x j There is a data point x in the first k nearest neighbors of i , then these two data points are called mutual k-nearest neighbors, and can be expressed as formula (4) and formula (5): M k (x i )={x i ∈N k (x j )^x j ∈N k (x i )} (4) Among them, N k (x j ) is x j The k nearest neighbor set, N k (x i ) is x i The k nearest neighbor set of Represents x i The nth mutual k-nearest neighbor of k (x i ) is an empty set, then prove that x i There are no mutual k-nearest neighbors, that is, at this time the data point x i The k nearest neighbors of x are all pseudo-nearest neighbors. i Belong to outliers or border points; Step 4: When the number of mutual k-nearest neighbors is 0 or 1, choose to use x when calculating the mean and standard deviation of the data. i The number of data points N in the neighborhood of the first nearest neighbor replaces the number of its mutual k-nearest neighbors K; here the formula for the data mean and standard deviation can be changed to formula (6), the neighborhood set of the first nearest neighbor of the training data point xi It can be expressed by formula (7): The standardization formula is updated to the first equation in equation (8). When the number of mutual k-nearest neighbors is greater than 1, the standard deviation used in the data standardization process is as small as possible, so the standard deviation of the first nearest neighbor set is still used here, which also reduces the calculation amount of the model to a certain extent. In order to ensure that the standardized data follows a normal distribution with a mean of 0, x is selected. i The mutual k-nearest neighbor set mean replaces the first nearest neighbor set mean, and the standardized formula is improved accordingly; at this time, the improved standardized formula is the second equation in formula (8); The standardized formula obtained through the above analysis is shown in formula (8): Where x i * is the standardized data, and are the neighborhood mean and standard deviation of the first nearest neighbor; m(Mk(xi)) is the value of x i The mean of the first K nearest neighbor sample sets, K represents x i The number of mutual k-nearest neighbors of ; A PCA process monitoring model is established based on the information related to the mutual k-nearest neighbor neighborhood; Step 5: Perform an orthogonal transformation on the standardized data and the transformation matrix is The data set is decomposed into the principal component subspace and residual space as shown in formula (9): X=TP T +X=TP T +TP T (9) Where T∈R m×l is the score of X in the l-dimensional principal component subspace, Residual, P∈R n×l is the loading matrix of the principal component subspace; To detect the process, two traditional statistics T 2 and SPE with their control limits and SPE α According to the following formula: T 2 =x new PΛ -1 P T x new (10) where θ r and h 0 Satisfying formula (14): Among them, Λ in formula (10) -1 is the diagonal matrix of l largest eigenvalues; I in formula (11) is the unit matrix; F in formula (12) α (l,nl) indicates the critical value of the F distribution when the significance level is α and the degrees of freedom are l, nl; Step 6: The fault detection phase consists of the offline modeling phase and the online modeling phase to complete the detection.

2. The industrial process fault detection method based on mutual k local neighborhood standardized analysis according to claim 1, It is characterized in that The offline modeling steps of step 6 offline modeling phase are as follows: 1) Use normal data of different modes as the training dataset X∈Rn×m; 2) For each training sample point x i , calculate the distance between it and other sample points in the training data set according to the Euclidean distance formula; sort them in ascending order, and filter out the training sample point x according to the sorting result i The first k nearest neighbor points of make 3) For the selected neighbor sample set N k (x i ) for each sample point Repeat step 2 and filter out The first k nearest neighbor sample points of 4) According to formula (4), select the training sample point x i The mutual k sample set M k (x i ); 5) If M k (x i ) is an empty set or contains only one training sample point, then calculate the training sample point x i The mean and standard deviation of the first nearest neighbor of x are calculated using equation 1 in (8) i To standardize; 6) If M k (x i ) contains K training sample points, and K>1, then calculate the training sample point x i The mean of the mutual k-nearest neighbor sample set, And use equation 2 in equation (8) to calculate x i To standardize; 7) For each standardized training sample point x j * The principal component analysis method is used to extract the principal components and perform dimension reduction; 8) Calculate the relevant T 2 and control limits of SPE.

3. The industrial process fault detection method based on mutual k local neighborhood standardized analysis according to claim 1, It is characterized in that The online modeling steps in step 6 of the online modeling phase are as follows: 1) When a new sample point x enters the data set new , calculate the Euclidean distance between the sample point and all sample points in the training data set, and sort them in ascending order, and filter out the sample point x according to the sorting result new The first k nearest neighbors N k (x new ); 2) In the modeling stage, the new samples are standardized using steps 3, 4, 5, and 6 in offline modeling. 3) Project the standardized new samples into the main residual space; 4) Use equations (10) and (11) to calculate new samples The statistic T 2 and SPE; 5) Comparison The statistic T 2 and control limits of SPE.